Feihong Yu and Jinshan Zhang, Steel defect detection based on multi-angle and multi-feature fusion network, Eur. J. Math. Appl. 4 (2024), Article ID 13.

Full Text: PDF

Volume 4 (2024), Article ID 13

https://doi.org/10.28919/ejma.2024.4.13

Published: 16/05/2024

Abstract:

With the rise of deep learning, various industries are adopting deep learning-based object detection algorithms to streamline general engineering operations. Particularly, a single-stage object detection algorithm striking a better balance between accuracy and speed has gained prominence. For the YOLO V5 algorithm, which is widely used in the industrial field, there are some problems in steel defect detection. This paper proposes some improvements to the YOLO V5 algorithm to make up for the shortcomings. Firstly, in the aspect of feature extraction, we propose a new building block for convolutional neural networks (CNN), namely Mres2Net, through the hierarchical feature fusion of single feature and the fusion of residual mechanism, the perception field of the network in the feature extraction part is increased. It makes the network more sensitive to the target. Second, in the aspect of feature fusion, this paper uses cross-fusion of adjacent scale features to fuse features. This method weakens the semantic difference between different scales in the fusion stage, making feature fusion more stable and less conflicting. Experimental results indicate a 6% accuracy improvement over traditional YOLO V5, with enhanced recall stability during training using the proposed method.

How to Cite:

Feihong Yu and Jinshan Zhang, Steel defect detection based on multi-angle and multi-feature fusion network, Eur. J. Math. Appl. 4 (2024), Article ID 13. https://doi.org/10.28919/ejma.2024.4.13